Video information in some video formats is interlaced. In one format, an image is displayed by painting every other line of pixels on a display, and then coming back and painting the intervening lines of pixels. For example, the odd scan lines are painted on the screen one by one, and then the even scan lines are painted. The entire image is called a frame. The first set of lines is a first field, and the second set of lines is a second field. In an NTSC video signal, for example, a frame includes 480 lines of pixels, where each line includes 720 pixels. Each field contains 240 lines of pixels, where each line is 720 pixels.
It is sometimes desired to increase the amount of video information by increasing the number of pixels in each field from 240 lines to a complete 480 lines. This process, called deinterlacing, doubles the amount of pixel information.
One technique for deinterlacing is pixel-based motion detection. See, for example, U.S. Pat. Nos. 6,166,773 and 5,473,383. Typically, pixel-based motion detection involves estimating the existence of motion for every single pixel of the field. These estimates are based on the value of the pixel. In real video, an object is typically represented by a large number of pixels. Movement of objects usually appears as changes in a large group of pixels between video fields. As a result, the change of a single pixel between fields often correlates to changes of its neighboring pixels. This correlation is, however, generally not exploited in pixel-based motion detection. Secondly, pixel-based motion detection generally uses information from only a few intensity values around the pixel for which motion is being estimated. Most noises are, however, burst noises and pixel-based motion detection fails in burst noise situations. Moreover, in real video a few pixels may happen to be the same between video fields when objects with repetitive textures move. Pixel-based motion detection may fail in such situations because pixel values do not change between successive video fields.
Another technique is adaptive diagonal interpolation, sometimes called directional adaptive interpolation. See, for example, U.S. Pat. No. 6,133,957. Directional adaptive interpolation calculates differences of pixel pairs, selects the pair with the smallest difference, and interpolates these pixels. Unfortunately, some real video includes pixel patterns where there is more than one pixel pair with the smallest difference. Moreover, the pair with the smallest difference may not generate the best interpolation result.
U.S. Pat. No. 5,410,356 describes another technique where motion compensation is used for deinterlacing. In this technique, an image field is divided into blocks. For every block of interest, a motion estimation engine finds a group of pixels that best matches the block in a defined searching range. A displacement vector or motion vector is typically predicted that describes the spatial translation from the block to the matched group of pixels. New lines of pixels are interpolated using the block of interest and the matched group of pixels. Such motion compensation techniques often do not do a job of predicting motion for use in deinterlacers. In real video, objects can rotate, turn and deform into other shapes. Objects can occlude each other partially or completely. Using simple spatial displacement to detect this type of motion often results in poor decisions. Additionally, using a displacement vector to interpolate often does not obtain adequately accurate results for deinterlacing applications. Object movement cannot always be represented as an integer displacement vector. An object can move a distance to a fraction of a pixel, and the group of pixels corresponding to a block can be different from the block. Finding an accurate match in such a scenario is often difficult without a more noise-tolerant motion detection scheme. In addition to these problems, motion compensation is computationally complex. The cost of realizing a deinterlacer using this technique may be undesirably high. Pixel information is moved from a field memory to the interpolation circuitry using line buffers. The required memory bandwidth of the pixel storing memory is also generally high, which further increases system cost because higher performance memories need to be employed.
An improved deinterlacing method is sought that can be efficiently realized in hardware.